Agents can't tell safe from dangerous

To an agent reading a UI, "Delete database" and "Download report" look equally clickable. Same rectangle, same cursor, same click. Nothing in the interface tells the agent that one action is harmless